Random multifractals: negative dimensions

and the resulting limitations of the

thermodynamic formalism
B. MANDEL13ROT
J3y J3E.\'OIT

Physics DepaTtment, JB.,ll T. J. Watson Research CenteT. YoTklown H eights.

New York 10598, U.S.A. and ~111athemalics DepaTlment, Y ale University.
1..\'ew liaren. Connecticut 06520. U.S.A.

Thc story describcd in this paper has startcd with the ·death or surv ival· criterion ,
which thc a uthor publis hed in 1972- 1974 and had obtaincd in 1968 ,,·hile
investigating Kolmogorov's hypothesis t hat the turbulent dissipation e(d x ) in a box
is log-normally distributed. Using this criterion. the present paper discusses the
concrete significancc of negatiYe fractal dimensions. They arisc in those random
multifractal measures, for which the Cramer function j(a) (the 'spectrum of
singularities ·¡ satisfics f(a) < O for certain val ues of a. It is shown that in that case
the strict · thcrmodynamical formalism' so le!y im·oh·es the form of j(a) in the rango
wherc j(a) > O, ancl conc~crns threc aspects of su eh measures: (a) the finc-grained
multifractal propertics, which are non-random and the samc for (almost) all
realizations: (b) t he propcrtics obtained by us ing thc 'partition functi on · formalism;
and (e) the 'typical ' coarsc-grained multifractal properties. Howcvcr. the j(a)s in t hc
range whcrc j(a) >O say nothing about t hc var iability of coarsc-grained propertics
bctwccn samples. A clcscription of these fluctuations. hence a fu llcr multifractal
description of the meas urc. is shown to be providccl by the ,-a]ues of j(a) in the range
where j(a) < O. We prefer to rescn·e the term 't hermodynamic' for the fine-grained
and partition-functional propcrties. and to say that the coarse-graincd properties go
beyond the thermodynamies. i.c. are not maeroseopic but 'mesoscopie ·.

Dedication
Whilc the material in this paper is severa! stagcs rcmo,·ed from anything done by
KolmogoroY. it is appropriatc that it should be dcdicated to t he mcmory of Andrei
Xikolaic\'itch. A token of hiR greatness is t hat his work spanned fie lds far removed
from onc a n othcr. I vow spccial admiration for his seYcral bricf forays in highly
specialized fields where he could not be more than a t ra nsient v isitor. His classic K41
thcory is of course t hc occasion of the prcscnt collcction of papers. His second foray
into turbulence. clase in contcnt to what I propase to \\Tite about, occurrcd in 1962
and was extremely bricf. Kolmogoro,· ·s (1962) papcr may even secm an expository
rcport on Obukhov (1962). but in fact puts forward a bold · third hypothesis · of exact
log- normality. Ancl carcful rcading by severa! authors has revealcd profound flaws in
this hypothcsis. W hcn a g ia nt stumblcs, it is safe to expect subtle issues to be
involvcd. \\'e must be grateful to Kolmogoro v for ha,·ing poin tcd out a path he did
not choosc to follow hi msclf Ycry far. and for spurring m u eh hard and rcwarding work.
Proc. R. Soc. Lond. A (199 1) 434, 79
Printed in Oreat Britain 79
80 B. B. Mandelbmt

l. Multifractal measures, Cramer renormalization, the geometric and

analytic forms of scaling, and the pitfalls of the partition function
(a) The distTibution of coaTse-grained multifmctal measuTes, and the issue of how the
pmbability densities should be TenoTmalized to ensuTe asymptotic collapse
T he purpose of this paper is to taekle an issue much more general than log-
normality. Xeverthcless, it is best to begin with the hypothesis of log-normality,
stated in t he vocabulary that is today curren t among physicists, and to use it to
describe sorne subtle 'anomalies ·. Assuming translational invariance, we denote by
e(dx) t he dissipation in a small spatial domain dx around a point x. The K41 t heory
postulates that e(dx) = edx 1 dx 2 dx 2 . The K62 theory starts with the postulate that
e(dx) is random. Being defined for domains dx, e(dx ) is a random measure.
It is customary in the study of multifractals to write e(dx ) = €1dx l". This notation
is used in my book (Mandelbrot 1982, p. 372) and in hundreds of more recent
references ; a(dx ) is a coarse-grained version of the Holder exponen t.
To specify a random function of t ime, one needs the probability djstributions of its
increments over different t ime increments dt . Similarly, the first thlngs to know
about a measure e(dx) are t he probability distributions for all ldxl of the value of €
(or of a) when the domain dx is choscn at random. Because of translational
invariance, the probability density of a can be written as p(a , dx). Statistical
mechanics and the d imensional analysis that underlies all fractal considerations
suggest t he following question. 'Can the p(a, dx), corresponding to differen t values
of ldxl be renormalized so th at they collapse into a single expression, at least for
ldxl -+ O? '
T he key fact about mult ifractal measures is that such collapse occurs. The
collapsed expression, a function of a, is usually denoted by f(a). For reasons that will
become clear in § 1 d, we shall call it CmmeT function. It describes the mult ifractal
properties of the measure e(dx).
Thc specific goal of this paper is to discuss cases where f(a) < O for sorne valu es of
a . This will complete a task I undertook many years ago .

(b) On self similaTity. or gemnetTic scaling. veTsus analytic scaling

Before we define j(a), it is useful to excerpt from the Manifesto in pmise of explicit
and visualized geometTy in Mandelbrot (1984). 'A theme t hat runs increasingly
strongly in my work is that explicit and visualized geometry is important in science
and in mathematics. Blind analytic manipulation is neveT enough. Formalisms,
however effective in the short run. aTe neveT enough. Many quantities that have
originated in geometry even t ually come to be used only in analytic relations; but to
forget geometry, and simply identify the fractal dimension D with an analytic
quantity. is not enough.'
One may add that the multifractal formalisms , when used mechanically, neglecting
their geometric meaning, a re not enough. The proof has been provided by t he very
fact that two schools of tho ugh t define the notion of multifmctal in profoundly
different ways.
\iVe must therefore tackle the very distasteful topic of definitions. Informally,
'fractal sets and multifractal measures are geometric objects such that each small
part is very much like a red uced size image ofthe whole'. Unfortunately. no single
formal definition fits, and I use the following realistic semi-formal description. 'A
fractal is a random or non -random set that is geometrically scaling. A multifractal
Proc. R . Soc. Lond. A (1991)
 Negative dimensions and limitations oj thermodynamics 81
is a random or non-random measure that is geometrically scaling. Either may be
linearly self similar or linear!y self affine, or may perhaps fail to fulfil either condition
exactly, but come close enough to be handled with the help ofthe same mathematical
tools.'
For othcr authors, however, ' fractals and multifractals are geometric objects that
satisfy di verse analytic scaling relations ' .
In thc case of fractals on which mass is distributed in fractally homogenous
fashion , the mass M(R) within a radius R satisfies M(R) ~ RD. There is competition
between a geometric description of fractal sets and an analytic description based on
the above M(R). But in practica! terms, this has not mattered much.
In the case of multifractals, the situation has proven to be surprisingly different.
Since analytic scaling is not directly linked to specific implementations, one may
have expccted it to be the more general notion. But in fact it has proven to be
narrowcr, and multifractal measures that are geometrically but not analytically
scaling pro ve to be needed to study two topics of fundamental character: turbulence
and fractal aggregates. This creates the need for a sharp distinction.
The broad multifractals will be defined as being geometrically scaling. The
simplest and best known are the multiplicatively generated multifractals, either non-
random (as studied by the school of A. S. Besicovitch at the University of
Cambridge) or random (as studied in my works listed in the References). These
measures are exactly renormalizable, by design. Diverse analytic scaling relations
hold in most cases, but other analytic scaling relations either fail to hold or are of
restricted validity.
The narrow multifractals will be defined as being analytically scaling in a strong
sense. Therefore, they exclude the broad multifractals for which analytic scaling
either fails or is weakened.
At one t ime, this distinction seemed to lack practica! bite, but it is proving
increasingly important. In particular, the present paper is devoted to two results. (a)
The requirement of analytic scaling excludes the random multiplicative multifractals
withj(a) <O. (b) Deep differences exist in the extent to which different multifractals
are represented by the 'thermodynamical formalism '.

(e) The partition Junction and its pitjalls

The most widely used analytic scaling relation concerning multifractals involves
the partition function x(q , ldxl) = :Eeq(ldxl). It concerns a measure e that was coarse-
grained by boxes of equal si de ldxl. Divided by a total number of boxes, X is a sample
qth moment of the coarse-grained measure e(dx).
Random or non-random measures constructed by a multiplicative cascade (see §2
below) resulting are geometrically scaling , and the expectation of x(q, ldxl) , if finite,
takes the analytically scaling form ldxlr(q) (Mandelbrot 1974a, b). In the non-random
case, 'expectation ' is of course to be replaced by actual value. When ¡.¿ > O for all
boxes and the ldx l are equal (as we shall suppose throughout) , this X is finite. The
revival of interest in multifractals first limited to the non-random case. The function
T(q) was reintroduced in Hentschel & Procaccia (1983), and Halsey et al. (1986) define
a (non-random) multifractal as a measure satisfying x(q , dx) = ldx¡r<a>.
When physicists moved from non-random to random multifractals, it seemed
reasonable to use t he same formalism , reasonable not to envision the possibility t hat
expectations can diverge, and reasonable to expect the sample X to follow the same
analytical scaling rule as its expectation. For these reasons, the analytic scaling of
the sample x is widely used to define the notion of multifractal.
Prrx. R. Soc. Lond. A (199 1)
82 B. B. Mandelbrol
Unfortunately. those apparently reasonable pseudo-ergod ic prcsuppositions were
ill-inspircd. The sample and population expressions turn out to be \'ery distinct
notions in thc contcxt ofrandom multifractals. The function 7(q) is finite in the non-
random case. and is necessarily an incrcasing function of q, for all qs. So is (Jet me
add) the ratio T(q)jq. In the random case, to thc contrary. the funct ion T(q) bascd
upon expectations may only be defin ed in sorne r estricted rangc, q~ 1 n < q < q~ax· and
7(q) and T(q)/q need not be increasing. T his differcncc is a clue to why probabilistically
interesting results in Mandclbrot (1974a. b) and Kahane &. Pcyricrc (1986) show that
many · physically reasonable · anticipations fail in the case ofmu ltifractals. There are
many random cases of central importance such that analytical scal ing holds for the
expectation (cq). whcre the expectation is carried O\'Cr the wholc ensemble or
population of es, but cither fails for thc samplc partition function x, or holds in a
wcakened form.
(d) C'ramer renormalization in the log-normal case , and the C'ramer limit
funrtion C'(a)
K olmogoro,· ·s hypothesis of log-normali ty makes no explicit refcrcnce to thc
mechan ism that generates the measurc c(dx ). It merely postulates that there is a
paramctcr p.> O, su eh that thc probability density p(a. dx) is t he gaussian
(kjen¡t)Íexp { - k[a- ( 1 + ~p)f / 2p.}.
The factor k in this cxpression stands for logbldxl. whcrc bis an intcgcr called basis,
to be cxpla incd shortly.
It can be secn (Mandelbrot 1974a, b, 1989) that this hypothcsis is untenable.
mcaning that a meas urc ha,·ing this p(a , dx ) cannot be implemcntcd. Thc nearest
thing eonccptually is a multiplicatiYe measurc that is constructed using log-normal
multipliers. but does not itself involve log-normal c(dx) . Howcvcr. we shall find it
uscful to list sorne strange and anomalous consequences of log-normality. Thcn we
proceed toa carefu l analysis that shows that one conscquence is indced irremediably
unaceeptablc. but another lcacls to a varicty of interesting faets.
As mentioned in § J a. thc basic issue conccrning multifractals is t hat of data
collapse. The abovc cxpression for p(a, dx) can indeed be renormalized in su eh a way
that it eeascs to depend as ldxl. This is s hown directly but only in passing in
~landelbrot (1974b. p. 357) and in detail but only implicitly in Frisch & Parisi (1985 )
ami H a lsey el al. (1986). Thc key is to perform. what I propose to call. Crame1·
renonnalization by considcring the exprcssion
C'(a.ldxl) = ln p(a , dx)/ln lclxl.
L'nder the hypothcsis of log-normality .
lim,dx .0 C'(a.lclxl) = - l- (1 + !P )] 2 j 2¡t = C'(a),
by definition of C(a). T hcreforc. Cramer renormalization yields asymptotic collapsc
for the log-normal probability distributions. This collapsc m ight haYe bccn a
sercndipitous conscqucncc of log-normality. In Lhe morE' general cases I ha ,-e studied
and shall summarizc in §:2, e is built by a recursi,·c multiplicativc cascade. and
Cramcr collapse results from a thcorem by Harold Cramer. In t hc arproach in Halscy
el al. (1986). a result that is a special case of Cramer collapse comes out of a
mechanicalmanipulation of symbols. In any CYCnt. it remains in each caseto deduce
('(a) from thc rules governing e. and to clraw consequcnccs.
Proc. R . Soc. L ond. A ( 199 1)
 Negative dimensions and limitations of thermodynamics 83

2. Multiplicative multifractals and the 'anomalies' of the log-normal

O(a) or j(a)
(a} Background: the binomial multijractals
The binomial measures depend upon a single real paramcter m 0 , variously callcd
a multiplicr or a mass. Onc assumes that O < m 0 < 1 a nd m 0 =1= !, and one defines
m 1 = 1- m 0 . The basic 'gcnerating step ' spreads mass o ver thc halves of every dyadic
intcrval, with t he relativo proportions m 0 and m 1 placed to the left and to the right.
Thus, the first stage y ields the mass m 0 in [0, !J and the mass m 1 in [!, 1). M ter k
stages, let c:p0 and cp 1 denote the relative frequencics of Os and l s in t he binary
development of t = 0.(11 (12 •.• (Jk written in the counting base b = 2. With the binomial
mcasurc, thc dyadic in ter val [dt] = [t,t+2k] reccivcs thc mass ,u(dt} = m~"'·m~"'' · It
follows that a = a(c:p0 , cp 1 ) = -c:p0 log 2 m0 -cp1 log 2 m 1•
Asto thc n umbcr of intcrvals leading to c:p0 and cp1 , it is N(k , c:p0 , cp1 } = k!/ (kcp0 )!(kcp 1 ) 1
Hence, thc similarity dimcnsion of the set where this a is obscrvcd is
o(k.c:p0 ,c:p 1 ) = - ln N(k,cp0 ,c:p1 )/ln (dt) = - In lk!/(k({Jo)!(kcp1 )!]/ln(dt}.
For largo k , Ji m k ,. 00 o(k, c:p0 , cp1 ) = -c:p0 log2 cp0 - cp1 log2 cp1 . Eliminating cp0 and cp 1
between a and o, we obtain a function j(a} written in parametric forro.
It is extrcmcly important to restate j(a} in terms of rescaled doubly logarithmi c
plots of probabili ty dcnsitics. Thc binomial measures are, of course, not random , but
a random variable appears when a dyadic box of length z- k is picked at random
among t he 2k boxes. Then the H older exponcnt bccome a random varia ble, to
be denotcd by H. The probability of hi tting a prcscribcd a, to be denoted by
Pr{H = a}, equals z- kN(k , c:p0 , cp1 ) and satisfics
O(k,cp0 ,c:p 1 ) = - lnPr{H = a}f ln dt = - ln z- kN(k,c:p0 ,c:p1 }/ln dt.
H ence, thc Cramcr 's function is O( a}= lim k...,.oo O( k, c:p0 ,c:p1 ) = j(a} - l. More gen-
erally, if e is bu ilt on a s upport other than the interval lO, 1], one has
j(a} = O(a} +dimension of the meas ure's support.
This rclation is of broad validity, and allows us to evaluate j(a) via the Cramer
function O(a}.
(b} R andom multiplicative measures
The salicnt fcature of the Besicovitch measurcs in § 2a is that each st age of
construction multiplics a m ass by a non-random factor m 0 or m 1 . T o generalize
(Mandclbrot 1974a , b). we allow this multiplying factor to be random . For example,
considcr t hc base b box of length b- 1 starting at t = 0.(1000 .... To each fJ with
O~ fJ ~ b- 1 corresponds a random multiplicr M((J).
By a repetit ion of this schcmc , the b-adi c box of length b- k starting at t =
0.(11 (12 ... fJ k is assigned t he mass
,uk(dt} = M((J1 )M((J1 , (J2 } ••• M((J 1 , .•• ,(Jk ) = TIM.
Here, the succcssivc random multipliers J.V.l for g iven tare independcnt and we shall
assume t hem to be id entically distrib utcd .
Plotting ,u(dt) in logarithmic scale and normalizing by d ivid ing through Iogb dt, and
writing -IogbM = V, our box of base b yields
H = ln,u(dt}/ln (dt) = (1/k)L- logbM((J1 )- 1ogbM((J1 , (J2 } ... ] = (1/k} LV.
Proc. R. Soc. Lond. A ( 199 1)
84 B. B. JYiandelbrot
Thus, the random variable H is simply the average of k independent random
variables.

Strict and average conservation of mass

To fit the image of 'mass ', it is natural to assume sorne form of conservation. In
a conservativo construction, the point of coordinates 111(¡3) is a random point in the
portian of b-dimensional space defined by M(¡J) ;<: O and L..M(¡J) = l. This implies
M~ 1 and (M) = b- 1 . A variant construction , called canonical (lVIandelbrot
1974a, b), assumes that theM(¡J) are statistically independent, without the constraint
M< 1, and that mass is only conserved on the average, meaning that (M)= b- 1 .
These distinctions are important, but it has been shown (Durrett & Liggett 1983)
that the rules of dependence between the multipliers do not change the death or
survival criterion. The fact that they do not affect the multifractal formalism may
even be viewed as demonstrating that this formalism treats as being different
measures that may otherwise be important.

(e) Standard limit theorems and the Cramer theorem

We denote by p(v) probability density of V and by Pk(v) the probability density of
L.. V. In general, the probability density kpk(ka) of H = (ljlc) L.. V is impossibly
complicated and its limit distribution for k-+ oo is necessarily in vestigated using
limit theorems . The plural is needed, beca use convergence of random variables can
be defined in a variety of very different ways, hence there is a multiplicity of limi t
theorems, all different and each true on its own terms.
The law of large numbers tells us that if (H) < oo then H converges to (H), which
implies that C(a) has its maximum for a 0 = (H). And the centrallimit theorem tells
us that if (H2 ) < oo then C(a) is parabolic in the immediate neighbourhood of a 0 =
(H). In the multifractal context, these results are without interest, the reason being
that they give little weight toas far from a 0 . To determine C(a) , one needs the very
different ' large deviations theorem' of Harold Cramer (see Chernofl' 1952 ; Daniels
1954) . The Cramer theory has been rather obscure until now, but it is likely to soon
beco me widely known and used. As k --'>- oo, the ' local Cramer theorem ' asserts that
(1/ k) Iogb (probabi!ity density of a) converges to a limit C(a) .

'l'he Legendre formalism

The Cramcr theory also shows thatj(a) = C(a)- 1 and r(q) = - 1-logb(Jlfq) are
linked by the Legendre and inverse Legendre transforms. It is natural that the best
proof ofCramer's result, in Daniels (1954), should use the steepest descent argument.

(d) Four anomalies of the lag-normal J(a), and the key lo their solution
The idea has spread that tho graph of j(a) always has the n shape characteristic
of the binomial, including the property that j(a) > O. In fact, the n
shape is only
en countered for multifractals that belong to a limited class I have called rostricted
or narrow. By this standard , the parabolic graph in § 1 d presents numorous
' anomalies '.

The a 1 <O anomaly

This is in appearance the most absurd of all four. Its background is that the
function J(a) satisfies j(a) ~a , and that the equation f(a) = a has a single root a 1 ,
Proc. R. Soc. Lond. A (1991)
 Segative dimensions and limitations oj thermodynamics 85
which for narrow multifractals is both the Holder exponent and the dimension of the
mcasurc t heoretical support of the measure ¡¿. In the log-normal case, howe,·er, onc
finds that a 1 = 1-t¡¿. It follows that a 1 < O when ¡¿ > 2: how can a Holder exponent
be negati,-d This feature is to be discussed in §3a.

The a < O anomaly

This is a second most absurd anomaly, and is encountered for all values of ¡¿,
evcn in cases "·hen the first anomaly is absent. Log-normality predicts t hat (e,·en if
¡¿ < 2 so that a 1 > O) a can take negative values. This feature implies t hat a need
not be a Holder exponen t. It raises very interesting points that are beyond the scopc
ofthis paper. Mandelbrot (1990c) shows that the function O(a) obtained in §1d only
holds do\Yn to a = acrit = -1 + tfl; for a < a crit• one must replace this expression by
O( a) = 2ajfi-1.
The j(a) <O anomaly
It is next in apparent absurdity. The log-normal allows j(a) to be negati,-e for
a> a~ax = 1 +t+ v(8¡¿) and a< a~n = 1 +t#- v(8fi}. This fact is observed
(without explanation or comment) in Frisch & Parisi (1985). Negativej(a)s also occur
in non-log-normal cases for which the a 1 < O and a <O anomalies are both absent.
This feature is to be discussed in §3b.

K ey lo the solution of the anornalies

I propase in this paper to explain the a 1 < O and j(a) < O anomalies, hence to
eliminate them as potential paradoxes. The argument consists in extending the
solution of t he a 1 <O anomaly as given in Mandelbrot (1974a,b). The key is a
surprising ·death-or-survival (nos) criterion ', namely in a mathematical theorem
that im·olves the following feature. vVhen X= limn - :x: Xn (with Xn ~ 0), it may very
well happen that <Xq) < limn.... :x: <X~> - A physicist would have expected equality to
hold in all cases, and would ha ve considered the known cases of inequality as being
'pathological counter examples '. The standard counter example given in math-
ematics textbooks is indeed pathological. It occurs when X n =O with probability
1-1/n. and Xn = n with probability n . Clearly, X= O and <Xq) =O for all q >O, but
<X~) = nq-l ; therefore <Xq) < lim<X~) for q > l. It so happens, however, that the
study of random multifractals is full of instances where the moment of a limit need
not be the limit of the moments. I n particular, a sample average or sum (such as the
partition function) may very well behave differently from the corresponcling
population expectation. When su ch anomalies are present, t he average and t he
expectation may obey entirely different scaling rules.

3. The death or survival criterion and thermodynamics

(a) 'Phe solution of the anornaly oj a 1 <O involves the death or survival criterion
To restate this first anomaly. When ¡¿ > 2, the quantity a 1 = 1-tfl satisfies
a 1 < O. How can it be ?
The ans,,·er in volves the fact that, irrespective of the ,-alue of ¡¿, there is no such
thing as a measure with a log-normal e(dx). The nearest thing is a measure obtained
by a multiplicative cascade whose multipliers are log-normal random Yariables.
When fl < 2. this cascade does indeed generate a non-degenerate measure. This
meas u re is said to 'surv ive '.
Proc. R. Soc. Lond. A (1991 )
 86 B. B. Mandelbrot
When fl > 2. to the contrary. the limit measure generated by lag-normal multiplicr
is identically zero for all ldxl. It is said to 'die· and a 1 is nota meaningful notion. The
caleulation that yields a 1 < O is just a formal exercise that cannot be made rigorous.
Since it does not evaluatc any meaningful quantity. the fact that its outcome is
negativo means nothing.
ln general random multiplicative multifractals. the nos criterion involves the
quantity a 1 = (MlogbM). When a 1 >O, the mcasurc survives; when a, <O, the
measure dies. vVhen the cascade preserves mass. it necessarily survives. because a 1 >
O and Le(dx). In all othcr cascades with a 1 > O. the sum Le(dx) converges to a limit
.random Yariable (Mandelbrot 1974a.b: Kahane & Peyriere 1976 ; Durrett & Liggett
1983). \\' hen thc cascade starts with measure l. the limit mass is currently denoted
(Mandelbrot 1990c) by the letter Q.

(b) To solve the j(a) <O anomaly, it su.ffices to apply the death or
survival criterion to the multifractal e<q>, which is defined as being constructed with the
multipliers M <q> = Mq jb(lllq) = "'"l1qb-'(q)
Indccd. taking ldx l = b-k. the expression x(q. dx) = L: ej?(dx ) is simply the product
by (bk)'(q) of the sample su m of the multifractal e<q> as defined in the title abo,-e is a
random ,-ariable and we know from §3a that its beha,·iour depends on the sign ofthe
quantity a~q> . defined as thc a 1 relative to the measure e<q>.
Beforc wc pursue, let us point out that aiq) > O when q is in thc range qri'un < q <
q~ax which corresponds j(a) > O. hence to a~in <a < a~ax; when j(a) < O, one has
a\q) <O. Indeed, define ¡<q>(a) as the function f(a) relativo to e<q>. To obtain
geomctrically the graph of j<q>(a). one takes the following stcps. (a) Expand a an<!..
j(a) horizontally in thc ratio q. obtaining a ncw Yariable a<q) and a new function
J<ql(a). (b) Trace the tangcnt of slope 1 of j<q>(a). This is the expanded form of the
tangcnt of slope q of f(a). (e) Reset the origin of a<q) at the point where the tangent
intersccts the axis/= O. This will define the quantity a<ql . \Ve can now apply the death
or surviYal criterion to M<q).
When a\q) <O, thc DOS criterion tells us the measure constructed using the
multipliers .LJ1<ql. i.e. the sequen ce of the quantities e<q>. converges toa non-degenerate
random limit. which is the ,-ariable Q (q) corresponding to the new multiplier J,f (q) .
Therefore. in agreement with general expcctation. the sample sum x(q. dx) can serve
to estímate the population expectation bk(ea(dx)). In particular, the sample sum and
the population expcctation follow precisely the same analytic scaling r ule, namely
x(q , dx) = bkT(q)Q(q). Thcrcfore. the sample r(q ) is idcntical to thc population r(q).
When a~q) < O, to thc contrary. the sample sum of e<q) converges to zero. The
uncxpected consequence is that geometric scaling of fl fails to imply that the sample
su m x(q. dx) is analytically scaling. Howevcr. an clcmentary but lengthy argument,
for whi ch we have no spacc. shows for q < q~in the sample sum is asymptotically
scaling for dx -+ O, with the scaling exponent f(q) = qa~in- whi le for q > q~ax· thc
asymptotic scaling cxponent is f(q) = qa~ax· The Legendre transform of this 'f(q)
happcns t'o be simply the portian of J(a) whcre j(a) > O. The portian where j(a) <O
plays no role whatsoever in the asymptotic study of X· However, this 7(q) is only
defined asymptotically for dx -+ O. This may explain the reports that j(a) < O has
been obtained from the partition function, and could not be aecountcd for.
\Ye ha,-e carried out cxtcnsi,·e computations to test this prediction; the results are
very striking. The rcader is adYised to test thcm again, for cxample using t he
measurcs described in l\Iandelbrot (1989b, 1990 a).
Proc. R. Sor. Lond. A (1991 )
 Negative dimensions and limitations of thermodynamics 87

(e) The notion oj thermodynamic Jormalism and its limilalions

In every approach to multifractals , the actual manipulations t hat linkf(a) to T(q )
are dircctly borrowed frorn thcrmodynamics. For example, the probabilists who
work with large deviations vicw thc Cramer t heory as being thermodynamical. Also,
whcn askod to introduce j(a) and 7(q) in the simplcst contcxt, I prefer to use the
Lagrange rnultiplicrs path towards the thermodynamics (l\'landelbrot 1989a , §6). As
to thc a pproach in Halscy et al. (1986), it uses stecpest dcsccnts, hcnce (without
saying so) uses the Darwin- Fowlcr approach of thermodynamics.
However , noticing analogies between formalism s is not cnough. As we dig deeper,
wc find ambigu ities in the seope that should be given to thc notion of thermodynam ic
propcrty.
One rnight have argued that it should cover everything that deals with
asyn1ptotics. In thc prcscnt case, the relevant asymptotics is the Cramer theory .
With t his definition , one would describe the whole function j(a) as thermodynamic,
including thc part in which f(a) <O. In my opinion, however, this view would be
incl usive to cxccss. Even though thc valucs of j(a) < O are obtained by a limit
thcorem , the actual cstimation of J(a) <O must rely upon methods that involve
prcasy mptotics; for example it can use the method based on actual histograms
within ' supersamples ' (Mandclbrot 1989a, §4.3- 4.5; 1990c, §5.8).
l wou ld more readily argue that the notion thermodynamic property should be
reserve to considerations that in volve the partition function. This restrictsj(a) t o t he
valucs J(a) >O.
Furtherrnorc, my earlier studies of J(a) <O (Mandelbrot 1989a, §4.3-4.5; 1990c,
§5.8) show that the portion J(a) >O can serve to define the notion of ' typical '
multifractal propcrtics in the case when they are gcnuinely random, which happens
to occ ur when dx is small but positive.
Thc samc studics can also be pushed (but we have no room hcrc for this task ) to
show that, as ldxl-+ O, the multifractal properties of e(dx) cease to be random, and
t hat thoy are fully determined by thc portion J(a) >O.
It remains to give a namc to the properties that depend on the portion whercf(a),
hcncc are not stri ctiy thermodynamical. I propose to call them mesoscopic.
There is no room hcrc to summarize the properties of multifractals with a left-sided
j (a) (scc, for example, M.andelbrot 1990b). They also involve mesoscopic effects,
which are of very different kind and of equally great practica! importance .
.Jilany useful discussions with C. J. G. Evertsz are acknowledged with pleasure.

References
Chernoff. H. 1952 A measure of asymptotic efficiency for tests of a hypothesis based on the sum
of observations. Ann. Math. Stat. 23 , 493- 507 .
Daniels. H. E. 1054 Saddlepoint approximations in statisti cs. Ann. Math. S tat. 25 , 631- 649.
Durrett. H. & Liggett, T. M. 1983 Fixed points of the smoothing transformation. Z. Wahr 64,
275- 30 1.
Frisch, U. & Parisi, G. 1985 Fully developed turbulence and intermittency. In Turbulence and
p1·edictability in geophysical fluid dynamics and climate dynamics (ed. M. Ghil), pp. 85- 86.
Amsterdam: Xorth - Holland.
Halsey, T. C.. J ensen, M. H .. Kadanoff, L. P. , Procaccia, l. & Shraiman, B. I. 1986 Fractal
measures ancl t hei r singularities: The characterization of strange sets. Phys. Rev. A 33 ,
ll4 l- l151.
Proc. R. Soc. Lond. A ( 1991 )
88 B. B. Mandelbrot
Hentschel , H. G. E. & Procaccia, I. 1983 The infinite number of generalized dimensions offractals
and strange attractors. Physica DS, 435-444.
Kahane , J. P. & Peyriere, J. 1976 Sur certa.ines martingales de B. Mandelbrot. Adv. Math. 22,
131 - 145.
Kolmogorov, A. N. 1962 A refinement of previous hypotheses concerning the local structure of
turbulence in a viscous incompressible fluid at high Reynolds number. J. FluidMech. 13, 82- 85.
Mandelbrot, B. B. 1972 Possible refinement of the lognormal hypothesis concerning the
distribution of energy dissipation in intermittent turbulence. In Statistical models and turbulence
(ed. M. Rosenblatt & C. Van Atta), p. 333- 351. New York: Springer-Verlag.
Mandelbrot, B. B. 1974a Intermittent turbulence in self similar cascades; divergence of high
moments and dimension of the carrier. J. Fluid Jlfech. 62, 331-358.
Mandelbrot, B. B. 1974b Multipli cations aléatoires itérées et distributions invariantes par
moyenne pondérée aléatoire. Comptes Rendus (Paris) 278A, 289-292 and 355-358.
Mandelbrot, B. B. 1982 The fractal geometry oj nature . New York: W. H . Freeman.
Mandelbrot, B. B. 1984 Fractals in physics: squig clusters, diffusions, fractal measures and the
unicity of fractal dimension. J. Stat. Phys. 34, 895-930.
Mandelbrot, B. B. 1989a Multifractal measures, especially for the geophysicist . Pure appl.
Geophys. 131, 5-42.
Mandelbrot, B. B. J.989b A class of multinomial multifractal measures with negative (latent)
values for the ' dimension' j(a). In Fractals' physical origin and properties (ed. L. Pietronero),
pp. 3-29. New York : Plenum .
Mandelbrot, B. B . 1990a Negative fractal dimensions and multifractals. Physica A 163 , 306-315.
Mandelbrot, B. B. 1990b New 'anomalous ' multiplicative multifractals: left-sided j(a) and the
modeling of DLA. Physica A 168, 95-111.
Mandelbrot, B. B. 1990 e Limit lognormal multifractal measures. In F1·ontiers of physics: Landau
memorial conje1·ence (ed. E. A. Gotsman, Y. Neéman & A. Voronel), pp. 309- 340. New York:
Pergamon. ·-
Mandelbrot, B. B. 1991 Fractals and multijractals: noise, turbulence and aggregates (selecta, vol. 1).
New York: Springer.
Obukhov , A . M. 1962 Sorne specific features of atmospheric turbulence. J. Fluid Mech. 13, 77-81.

Proc. R . Soc. Lond. A (199 1)